A baker follows a recipe for a cake, but instead of a fluffy sponge, they get a hard crust. They start running through possible causes: maybe the flour was damp, or the oven heats unevenly, or the eggs were too cold. That's roughly how calculations in materials science behave: the standard theory (DFT) often errs, predicting that a crystal will conduct electricity, while the experiment shows it's an insulator. Each such miss is a signal that something is missing from the model: magnetic order, strong electron interactions, a hidden defect.
Previously, finding the cause was a manual, time-consuming process, like a baker testing hypotheses for weeks. The new tool XDFT acts like a smart kitchen assistant: it compiles a list of suspects from a catalog of typical overlooked effects, runs trial calculations for each hypothesis, and sees which one best matches reality. Along the way, it learns on the fly — updating the Bayesian probability that a given class of hypotheses is useful at all. On a test set of 124 materials, the algorithm found the correct explanation for 70 out of 90 discrepancies. That's 78%, whereas random guessing gave only 19%, and a static list from a language model gave 20%.
In the end, for each material, XDFT provides not only a corrected recipe for the calculation but also a clear indication of which physical mechanism was missed. Cases where no explanation could be found are flagged as candidates for new experiments — perhaps something is hidden there that isn't in the catalog. Interestingly, all successfully solved cases could be sorted into just three groups of chemical elements that make up the substance. This allowed a short rule of four lines to be formulated — a kind of cheat sheet for quick diagnostics without heavy computation.
So the failed baker becomes a head chef who not only bakes perfect cakes but also understands why the dough didn't rise last time. And if a new, never-before-seen ingredient appears in the kitchen, they already know where to start checking.
Density functional theory (DFT) is one of the most cited methods in physics, but for many modern materials it systematically errs, predicting a metal where there is actually an insulator.
In Neal Stephenson's novel 'Anathem', scientists use 'syntactic' methods for automatic analysis of theories — somewhat similar to how XDFT iterates over hypotheses about material structure.